HMACE: Multi-Agent LLM Framework for Combinatorial Optimization

Researchers have unveiled HMACE, a Heterogeneous Multi-Agent Collaborative Evolution framework that reinterprets heuristic search as a design challenge for NP-hard combinatorial optimization. In contrast to current LLM-based approaches that depend on inflexible workflows with set templates, HMACE breaks down each evolutionary generation into a cycle involving four distinct agents: a Proposer for exploring strategies, a Generator for creating executable heuristics, an Evaluator for empirical evaluation, and a Reflector for updating memory with archived data. This framework integrates behavior-aware retrieval and lightweight candidate filtering, among other techniques, to prevent early convergence on local optima. The paper can be found on arXiv with ID 2605.07214.